Multi-agent systems that do not need babysitting

Agent frameworks are very good at showing you a task that completes. They are quieter about the run that loops for forty minutes, burns through a month's budget, and produces nothing.

Give every run a budget

Not a suggestion — a ceiling, enforced in code. Two of them:

• a token budget for the whole run, not per call
• a wall-clock timeout, because a stuck tool call does not consume tokens while it hangs

When a run hits either ceiling it should stop and report what it had, not silently keep going. This one change turns a scary open-ended cost into a line item you can reason about.

Log the trace, not the answer

When an agent produces something wrong, the final message tells you almost nothing. You need the full trace: every tool call, its arguments, what came back, and the reasoning between steps.

Store it. Make it searchable. The first week after launch you will read traces daily, and you will find that most failures are not model failures — they are a tool returning an error string the model then treats as data.

Fewer agents, clearer edges

A swarm of nine specialised agents demos beautifully and debugs terribly. Every handoff is a place where context is lost and instructions are re-interpreted.

Start with two or three, each with a boundary you can state in a sentence. Add another only when you can point at a concrete failure the split would have prevented.

Decide what "done" means

The most common cause of a run that never ends is that nothing defined success. "Research the topic thoroughly" has no terminating condition. "Return five sources with publication dates, or say which you could not find" does.

Write the stopping condition into the task, not into the system prompt where it becomes a suggestion.